Changelog

What the current DispatchAtlas tree provides, grouped by area. The canonical, machine-readable history lives in the repository CHANGELOG.md, which follows Keep a Changelog and semantic versioning once package releases begin; tagged versions appear on the releases page.

DispatchAtlas is a reproducible benchmarking platform for scheduling and optimization across distributed-computing environments — from classic combinatorial families to the IoE / Edge–Fog–Cloud continuum.

✨ Current Highlights

• Core scheduling model — unrelated-machine cost models (execution-time matrix, compatibility mask, sequence-dependent setup, load-dependent execution curves, cross-resource communication), multi-resource co-allocation, partial-execution (imprecise computation), moldable variable-width tasks, gang co-scheduling, and multi-tenant fair-share task semantics, a broad objective family (makespan, setup, energy, cost, carbon, latency, lateness, fairness, dominant resource share, reliability, security, robustness, imprecise reward, composite) with scalar-or-vector values, feasibility reporting with infeasible-row records, deterministic seed derivation, and JSON serialization behind typed integration protocols.
• Benchmark atlas — generic combinatorial-optimization scheduling families (unrelated-machine, job-shop, flexible job-shop, permutation flow-shop, sequence-dependent setup flow-shop, and resource-constrained project scheduling) alongside distributed-computing families and a continuum of Edge–Fog–Cloud families (edge offloading, serverless cold-start, microservice call graphs, replica placement, aerial edge, streaming windows, failure recovery, cyber-physical scheduling, IoT sensor-to-edge aggregation, GPU and machine-learning cluster scheduling, edge service placement, carbon-aware deferral, federated-learning client selection, immersive extended-reality rendering, low-earth-orbit satellite edge scheduling, datacenter task colocation, virtual-machine allocation, vehicular edge offloading, industrial time-sensitive networking, 5G network slicing, digital-twin edge synchronization, smart-city video analytics, collaborative edge-cloud split inference, confidential edge computing, battery-free intermittent edge computing, preemptible spot-VM scheduling, data-locality placement, coflow co-completion scheduling, and generative-model inference serving under a key-value-cache budget), each with a profile class, citation-backed metadata, characterization metrics, difficulty stratification, and a distribution-distance calibration metric.
• Solver registry — constructive and dispatching baselines, exact adapters (CP-SAT, mixed-integer linear programming, branch-and-bound, two native bounded solvers — exhaustive enumeration and decision-diagram sequencing — and optional commercial backends behind a dedicated extra; a quadratically-constrained backend stays registered in the internal-only pool), representative metaheuristics, a diversified competitor set with canonical references, continuous solvers bound to scheduling through encoding adapters, a many-objective competitor, and a native-encoding solver family that searches directly over feasible schedules. Every descriptor carries a fail-closed citation contract and declared capabilities.
• Selection layer — rule-based and learned solver recommendation that ranks solvers from declared capabilities and benchmark characterization, behind a leakage-free cross-validation protocol that reports held-out generalization and never claims a globally best solver.
• Objectives, constraints, and quality — per-objective evaluation, named service-level constraints with distinct hard-breach and soft-penalty paths, Pareto-frontier helpers and four quality indicators (hypervolume, IGD+, the additive epsilon-indicator, and spread), and worst-case or conditional value-at-risk robustness over an explicitly declared perturbation set.
• Campaign engine — validation, dry-run budgeting, checkpointed and replayable execution, a run-count policy at the statistical-power floor, fair-comparison invariants, deterministic replay against recorded content hashes, and a safe parallel topology that scales to the host's available cores with capacity probing and nested-parallelism guards.
• Analysis and evidence — non-parametric inference (Wilcoxon signed-rank with Vargha-Delaney and Cliff's-delta effect sizes, the Friedman omnibus with a Nemenyi post-hoc, Holm and Hochberg correction, and bias-corrected accelerated bootstrap intervals) with no third-party statistics dependency, deterministic SVG and PGFPlots figures and tables, a limitations surface for unsupported methods, and disclosure-filtered evidence bundles keyed by generic evidence tiers.
• Documentation and portal — a documentation site with an interactive portal (benchmark catalog explorer, results explorer, simulation preview, an explainable metadata-only solver recommender, platform inspection, and a download center), in-browser runnable demos (visualize evidence including convergence and exploration–exploitation figures, run a solver live, and run a seeded experiment with an effect size and a bootstrap confidence interval), offline search, a sitemap, a machine-readable site index, and public-safe portal data rendered into the page at build time.
• Governance and supply chain — root governance and community surfaces, REUSE-compliant single-line SPDX licensing, coverage-enforced CI, pinned dependency auditing, a scheduled supply-chain audit over the installed distribution closure, a CodeQL security configuration, CycloneDX SBOM generation, and release-signature and build-provenance verification.